CIPER: Combining Invariant and Equivariant Representations Using Contrastive and Predictive Learning